diff --git a/monai/transforms/utility/array.py b/monai/transforms/utility/array.py
index 297b243ff91..03dca190549 100644
--- a/monai/transforms/utility/array.py
+++ b/monai/transforms/utility/array.py
@@ -30,7 +30,7 @@
 from monai.config import DtypeLike
 from monai.config.type_definitions import NdarrayOrTensor
 from monai.data.meta_obj import get_track_meta
-from monai.data.meta_tensor import MetaTensor, _normalize_spatial_ndim, get_spatial_ndim
+from monai.data.meta_tensor import MetaTensor, _has_explicit_no_channel, _normalize_spatial_ndim, get_spatial_ndim
 from monai.data.utils import is_no_channel, no_collation, orientation_ras_lps
 from monai.networks.layers.simplelayers import (
     ApplyFilter,
@@ -231,6 +231,17 @@ def __call__(self, img: torch.Tensor, meta_dict: Mapping | None = None) -> torch
         else:
             result = moveaxis(img, int(channel_dim), 0)  # type: ignore
 
+        # When a channel dim is added (channel_dim="no_channel"), the affine stays
+        # the same but the tensor gains a channel axis, so the cached `spatial_ndim`
+        # (derived from the input tensor's shape) would be too low. Refresh it from
+        # the affine, which is the source of truth for spatial dimensions, so that
+        # downstream lazy transforms (e.g. spatial crop) build translations with the
+        # correct number of axes and preserve the affine offset in world space.
+        if is_no_channel(channel_dim) and isinstance(result, MetaTensor):
+            result.spatial_ndim = _normalize_spatial_ndim(
+                result.affine.shape[-1] - 1, result.ndim, no_channel=_has_explicit_no_channel(result.meta)
+            )
+
         return convert_to_tensor(result, track_meta=get_track_meta())  # type: ignore
 
 
diff --git a/tests/transforms/test_ensure_channel_first.py b/tests/transforms/test_ensure_channel_first.py
index fe046a4cdf9..9574aad1477 100644
--- a/tests/transforms/test_ensure_channel_first.py
+++ b/tests/transforms/test_ensure_channel_first.py
@@ -22,7 +22,7 @@
 from PIL import Image
 
 from monai.data.meta_tensor import MetaTensor
-from monai.transforms import EnsureChannelFirst, LoadImage
+from monai.transforms import EnsureChannelFirst, LoadImage, SpatialCrop
 from monai.utils import optional_import
 
 itk, has_itk = optional_import("itk", allow_namespace_pkg=True)
@@ -108,6 +108,23 @@ def test_default_channel_first(self):
 
         self.assertEqual(result.shape, (1, 4, 4))
 
+    def test_ensure_channel_first_no_channel_affine_regression(self):
+        # Regression test for https://github.com/Project-MONAI/MONAI/issues/9184.
+        # EnsureChannelFirst(channel_dim="no_channel") must refresh the cached
+        # `spatial_ndim` from the affine; otherwise a following spatial crop only
+        # shifts the first (N-1) axes and silently drops the last affine offset.
+        affine = np.eye(4)
+        affine[:3, 3] = [10, 20, 30]
+        img = MetaTensor(np.zeros((40, 40, 40), dtype=np.float32), affine=affine)
+
+        result = EnsureChannelFirst(channel_dim="no_channel")(img)
+        self.assertEqual(result.spatial_ndim, 3)
+        self.assertEqual(result.ndim, 4)
+
+        cropped = SpatialCrop(roi_start=[5, 5, 5], roi_end=[20, 20, 20])(result)
+        self.assertEqual(cropped.spatial_ndim, 3)
+        self.assertEqual(cropped.affine[:3, 3].tolist(), [15, 25, 35])
+
 
 if __name__ == "__main__":
     unittest.main()